I like Deep Learning concepts and their potential. However the problem is it is not scalable on budget free computation configurations. If we look at all the success stories about Deep Learning, either they devoted so much infrastructure or they run the code for weeks on lower budget hardware. Even the papers proposed in the field of Deep Learning mostly working on little size images, if they are not capable of those high budget systems.
I suffer from infrastructure and also I do not want to spend my time on 32x32 images, even my supervisor doesn't. Are there any implementations of Deep Learning that can work comparable with hand crafted methods on normal sources, or if there aren't then how can I add some deep learning ingredients to my image recognition problem whilst not being overwhelmed by the running time?
I think you mix two different things:
1/ the size of images which you use in the learning process,
2/ the size of images where you use the learner.
For example, to detect face images, you can learn on face and non face images of size 15x20. When you use the face detector on an image of size 1024x1024, you make a convolution of the face detector at different scales in order to find face images of different sizes.
There are some methods to speed up the convolution. See for example:
https://www.researchgate.net/publication/3193210_A_Fast_and_Accurate_Face_Detector_Based_on_Neural_Networks
Article A Fast and Accurate Face Detector Based on Neural Networks
But the current Deep Learning field is working on 32x32 CIFAR images and application to natural images at least in ~400x400 sizes are not very prevalent because of the reason I pointed out above bay. In addition resizing a natural scene image into 32x32 and applying Deep Learning to those would not give good results since you lose so much detail in that sizes.
A 32x32 images are sufficient to recognize a ferrari in a set of sport cars. This scale is sufficient for most of applications such as images / video annotation. When your algorithm works well on tiny images, it works well also on large images. The reverse is wrong.
Now if your application is face recognition for bank authentification, may be you need larger images. Does your application necessitate to learn (not to process) large images ?
Than assume you have a collection of images labelled like RED and BEACH and you aim to merge a model for recognizing these attributes over novel images. In your settings I need to downsize the image into 32 to 32 and come up a final filter that captures those labels over the overall image. Although it might work for a label like Beach since it is a concept sparked from the overall structure of the image, for a label like RED this overlook might be misleading since some little part of the image might be RED and that label is assigned by its confidence. Another problem is RED partition of the image can be so varying. It might be at the top of the image or at the very right of the image so it is not viable to model those very varying attributes from little images, if we aim to get a model at the end that is able to capture those concepts over novel images. At the other side, if we look at the CIFAR images the labelled concept generally at the center of the image and it is easy to capture even you downsize the image into 32x32. That is, what I mean is, for a general solution over all kind of labels and images (that is the ultimate purpose of the AI) we cannot assume that all the information is canned into 32x32 visuals, thus we need to be able to work on larger size images to capture more detailed attributes and concept. This is why I say that deep learning stacking into those little size images whereas the larger mediums waiting for certain recognition problems.
Eren,
I would have to agree with Raphaël. Also, if you need to process high resolution images I would also suggest first preprocessing instead of scaling the images, then passing them to the deep learner.
I wish you success in your endeavor
- Badges
- Science topic